Fidelitous Augmentation of Human Accelerometric Data for Deep Learning

Time series (TS) data have consistently been in short supply but needed to train systems in prediction, modeling, classification etc. Synthesis can serve to expand the sample population, yet it is crucial to maintain the statistical characteristics between the synthesized and the original TS to ensure consistent sampling of data. In such synthesis the time domain characteristics of the data may not be preserved which motivates for our work, the objective which is to retain for a synthesized TS: fundamental statistical attributes as well as important time domain features such as its overall shape and significant transients. In a novel way, we isolate important TS features into various components leaving the residual signal to be randomized. Using accelerometer data in a clinical setting, we compare our method to others and show it has higher fidelity to the original signal features, has good diversity and performs better data classification.